Está en la página 1de 4

A simple model of price formation

K. SznajdWeron , R. Weron

Institute of Theoretical Physics, University of Wroclaw,


pl. Maxa Borna 9, 50-204 Wroclaw, Poland
Hugo Steinhaus Center, Wroclaw University of Technology,
Wyspianskiego 27, 50-370 Wroclaw, Poland

arXiv:cond-mat/0101001 v2 7 Nov 2001

A simple Ising spin model which can describe the mechanism of price formation in nancial markets
is proposed. In contrast to other agent-based models, the inuence does not ow inward from the
surrounding neighbors to the center site, but spreads outward from the center to the neighbors. The
model thus describes the spread of opinions among traders. It is shown via standard Monte Carlo
simulations that very simple rules lead to dynamics that duplicate those of asset prices.
PACS numbers: 05.45.Tp, 05.50.+q, 87.23.Ge, 89.90.+n

generalized hyperbolic [13,14], and some reject any single


distribution [15,16].
Instead of looking at the central part of the distribution, an alternative way is to look at the tails. Most
types of distributions can be classified into three categories: 1 thin-tailed for which all moments exist and
whose density function decays exponentially in the tails;
2 fat-tailed whose density function decays in a powerlaw fashion; 3 bounded which have no tails.
Virtually all quantitative analysts suggest that asset
returns fall into the second category. If we plot returns
against time we can notice many more outlying (away
from the mean) observations than for white noise. This
phenomenon can be seen even better on normal probability plots, where the cumulative distribution function
(CDF) is drawn on the scale of the cumulative Gaussian distribution function. Normal distributions have the
form of a straight line in this representation, which is approximately the case for the distribution of weekly or
monthly returns. However, distributions of daily and
higher-frequency returns are distinctly fat-tailed [17].
This can be easily seen in the top panels of Fig. 1, where
daily returns of the DJIA index for the period Jan. 2nd,
1990 Dec. 30th, 1999, are presented.
Clustering and dependence. Despite the wishes of
many researchers asset returns cannot be modeled adequately by series of iid (independent and identically
distributed) realizations of a random variable described
by a certain fat-tailed distribution. This is caused by
the fact that financial time series depend on the evolution of a large number of strongly interacting systems and belong to the class of complex evolving systems. As a result, if we plot returns against time
we can observe the non-stationarity (heteroscedasticity) of the process in the form of clusters, i.e. periods during which the volatility (measured by standard
deviation or the equivalent l1 norm [18]) of the process is much higher than usual, see the top-left panel
of Fig. 1. Thus it is natural to expect dependence
in asset returns. Fortunately, there are many methods to quantify dependence. The direct method consists in plotting the autocorrelation function: acf (r, k) =
PN
PN
)(rtk r)/ t=1 (rt r)2 , where N is the
t=k+1 (rt r

The Ising spin system is one of the most frequently


used models of statistical mechanics. Its simplicity (binary variables) makes it appealing to researchers from
other branches of science including biology [1], sociology
[2] and economy [38]. It is rather obvious that Isingtype models cannot explain origins of very complicated
phenomena observed in complex systems. However, it
is believed that these kind of models can describe some
universal behavior.
Recently, an Ising spin model which can describe the
mechanism of making a decision in a closed community
was proposed [9]. In spite of simple rules the model exhibited complicated dynamics in one and two [10] dimensions. In contrast to usual majority rules [11], in
this model the influence was spreading outward from the
center. This idea seemed appealing and we adapted it
to model financial markets. We introduced new dynamic
rules describing the behavior of two types of market players: trend followers and fundamentalists. The obtained
results were astonishing the properties of simulated
price trajectories duplicated those of analyzed historic
data sets. Three simple rules led to a fat-tailed distribution of returns, long-term dependence in volatility and
no dependence in returns themselves.
We strongly believe that this simple and parameter free
model is a good first approximation of a number of real
financial markets. But before we introduce our model we
review some of the stylized facts about price formation
in the financial markets.
Stylized facts. Adequate analysis of financial
data relies on an explicit definition of the variables under study. Among others these include the price and the
change of price. In our studies the price xt is the daily
closing price for a given asset. The change of price rt
at time t is defined as rt = log xt+1 log xt . In fact,
this is the change of the logarithmic price and is often
referred to as return. The change of price, rather than
the price itself, is the variable of interest for traders (and
researchers as well).
Fat-tailed distribution of returns. The variety of
opinions about the distributions of asset returns and their
generating processes is wide. Some authors claim the
distributions to be close to Paretian stable [12], some to
1

PN
sample length and r = N1 t=1 rt , for different time lags
k. For most financial data autocorrelation of returns dies
out (or more precisely: falls into the confidence interval
of Gaussian random walk) after at most a few days and
long-term autocorrelations are found only for squared or
absolute value of returns [1820], see bottom panels of
Fig. 1. Recall that for Brownian motion the classical
model of price fluctuations [19,21] autocorrelations of
rt , rt2 and |rt | are not significant for lags greater or equal
to one.

fluctuation, respectively) and plotted against box size on


a double-logarithmic paper. Linear regression yields the
Hurst exponent H, whose value can lie in one of the three
regimes: 1 H > 0.5 persistent time series (strict long
memory), 2 H = 0.5 random walk or a short-memory
process, 3 H < 0.5 anti-persistent (or mean-reverting)
time series. Unfortunately, no asymptotic distribution
theory has been derived for the R/S and DFA statistics
so far. However, it is possible to estimate confidence intervals based on Monte Carlo simulations [23].
The model. Recently a simple model for opinion
evolution in a closed community was proposed [9,10]. In
this model (called USDF) the community is represented
by a horizontal chain of Ising spins which are either up or
down. A pair of parallel neighbors forces its two neighbors to have the same orientation (in random sequential
updating), while for an antiparallel pair, the left neighbor takes the orientation of the right part of the pair,
and the right neighbor follows the left part of the pair.
In contrast to usual majority rules [11], in the USDF
model the influence does not flow inward from the surrounding neighbors to the center site, but spreads outward from the center to the neighbors. The model thus
describes the spread of opinions. The dynamic rules lead
to three steady states: two ferromagnetic (all spins up or
all spins down) and one antiferromagnetic (an up-spin is
followed by a down-spin, which is again followed by an
up-spin, etc.).
In this paper we modify the model to simulate price
formation in a financial market. The spins are interpreted
as market participants attitude. An up-spin (Si = 1)
represents a trader who is bullish and places buy orders,
whereas a down-spin (Si = 1) represents a trader who
is bearish and places sell orders. In our model the first
dynamic rule of the USDF model remains unchanged,
i.e. if Si Si+1 = 1 then Si1 and Si+2 take the direction
of the pair (i,i+1). This can be justified by the fact
that a lot of market participants are trend followers and
place their orders on the basis of a local gurus opinion.
However, the second dynamic rule of the USDF model
has to be changed to incorporate the fact that the absence
of a local guru (two neighboring spins are in different
directions) causes market participants to act randomly
rather than make the opposite decision to his neighbor:
if Si Si+1 = 1 then Si1 and Si+2 take one of the two
directions at random.
Such a model has two stable states (both ferromagnetic), which is not very realistic for a financial market.
Fortunately, trend followers are not the only participants
of the market [27]. There are also fundamentalists players that know much more about the system and have a
strategy (or perhaps we should call them rationalists).
To make things simple, in our model we introduce one
fundamentalist. He knows exactly what is the current
difference between demand and supply in the whole system. If supply is greater than demand he places buy
orders, if lower sell orders.
It is not clear a priori how to define the price in a

0.08
0.999
0.99

Returns
0.04

CDF

0.90
0.50

0.10
0.04
0.08

0.01
0.001
500

1000 1500 2000 2500


Days

0.05

0.05

1
ACF

ACF

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0
0.2
0

0
Returns

0
10

20
30
40
Time lag (days)

50

0.2
0

10

20
30
40
Time lag (days)

50

FIG. 1. Daily returns of the Dow Jones Industrial Average


index during the last decade (top left), normal probability plot
of DJIA returns (top right), lagged autocorrelation function
of DJIA daily returns (bottom left), and lagged autocorrelation function of absolute value of DJIA daily returns (bottom
right). Dashed horizontal lines represent the 95% condence
interval of a Gaussian random walk.

Another way to examine the dependence structure is


the power spectrum analysis, also known as the frequency domain analysis. One of the most often used
techniques was proposed by Geweke and Porter-Hudak
[22] (GPH) and is based on observations of the slope of
the spectral density function of a fractionally integrated
series around the angular frequency = 0. A simple
linear regression of the periodogram In (a sample analogue of the spectral density) at low Fourier frequencies
k : log{In (k )} = a dlog{4 sin2 (k /2)} + k yields the
differencing parameter d = H 0.5 through the rela The GPH estimate of the Hurst exponent
tion d = d.
H has well known asymptotic properties and allows for
construction of confidence intervals [22,23].
Yet another method is the Hurst R/S analysis [24,25]
or its younger sister the Detrended Fluctuation Analysis (DFA) [26]. Both methods are based on a similar
algorithm, which begins with dividing the time series
(of returns) into boxes of equal length and normalizing
the data in each box by subtracting the sample mean
(R/S) or a linear trend (DFA). Next some sort of volatility statistics is calculated (rescaled range or mean square

The returns rt are obtained from the simulated price


curve xt (see Fig. 2) after it is shifted (incremented by
one) to make it positive. Alternatively we could have defined the up-spin to be equal to two and the down-spin
to zero. However, this would have made the calculations more difficult and the description of the model less
appealing.
In Figure 3 we present daily returns and normal probability plots for the simulated (left panels) and USD/DEM
exchange rate (right panels) time series. Without prior
knowledge as to the magnitude of historical returns it is
impossible to judge which process is real and which is a
fraud. The same is true for the simulated and the DJIA
returns of Fig. 1.

market. The only obvious requirement is, that the price


should go up, when there is more demand than supply,
and vice versa. For simplicity, we define the price xt in
our model as the normalized differencePbetween demand
N
and supply (magnetization): xt = N1 i=1 Si (t), where
N is the system size. Note that xt [1, 1], so |xt | can be
treated as probability. Now we can formulate the third
rule of our model: the fundamentalist will buy (i.e. take
value 1) at time t with probability |xt | if xt < 0 and sell
(i.e. take value 1) with probability |xt | if xt > 0.
The third rule means that if the system will be close
to the stable state all up, the fundamentalist will place
sell orders with probability close to one (in the limiting
state exactly with probability one) and start to reverse
the system. So the price will start to fall. On the contrary, when r will be close to 1, the fundamentalist will
place buy orders (take the value 1) and the price will
start to grow. This means that ferromagnetic states will
not be stable states anymore.
Results. To investigate our model we perform a
standard Monte Carlo simulation with random updating.
We consider a chain of N Ising spins with free boundary
conditions. We were usually taking N = 1000, but we
have done simulations for N = 10000 as well. We start
from a totally random initial state, i.e. to each site of
the chain we assign an arrow with a randomly chosen
direction: up or down (Ising spin).

0.1

0.04
Returns

Returns

0.05

0.02

0.05

0.02

0.1

0.999
0.99

500

1000 1500
Days

2000

0.04

0.999
0.99

CDF

0.90

0.90

0.50

0.50

0.75

0.10

0.10

0.9

0.7

0.01
0.001

0.01
0.001

0.8

0.65

0.7

0.6

0.6
500

1000 1500
Days

USD/DEM
2000

0.5

500

1000 1500
Days

0
Returns

0.05

1000 1500
Days

2000

CDF

0.02

0
Returns

0.02

FIG. 3. Returns of the simulated price process xt and


daily returns of the USD/DEM exchange rate during the last
decade, respectively (top panels). Normal probability plots of
rt and USD/DEM returns, respectively, clearly showing the
fat tails of price returns distributions (bottom panels).

0.55
Simulation

0.5

0.05

500

2000

FIG. 2. A typical path of the simulated price process xt


and the USD/DEM exchange rate, respectively.

In Figure 4 we present the lagged autocorrelation study


for the simulated (left panels) and USD/DEM exchange
rate (right panels) time series. Again the properties of
the simulated price process duplicate those of historical
data. The lagged autocorrelation of returns falls into
the confidence interval of Gaussian random walk immediately, like for the dollar-mark exchange rate. The same
plot for the DJIA returns shows a bit more dependence.
This can be seen also in Table 1, where values and
significance of the R/S, DFA and GPH statistics for the
simulated and four historical data sets are presented. In
all but one case (DFA for DJIA) the Hurst exponents
of daily returns are insignificantly different from those
of white noise. On the other hand, the Hurst exponents
of the absolute value of daily returns are persistent in
all cases, with the results being significant even at the
two-sided 99% level. Moreover, the estimates of H from
the simulation are indistinguishable from those of real
market data.

In our simulations each Monte Carlo step (MCS) represents one trading hour; eight steps constitute one trading day. We typically simulate 20000 MCSs, which corresponds to 2500 trading days or roughly 10 years. We
chose such a period of time, because in this paper we compare the empirical results with historical data sets (two
FX rates and two stock indices) of about the same size:
2245 daily quotations of the dollar-mark (USD/DEM) exchange rate for the period Aug. 9th, 1990 Aug. 20th,
1999 (see Fig. 2 and Table 1), 2809 daily quotations
of the yen-dollar (JPY/USD) exchange rate for the period Jan. 2nd, 1990 Feb. 28th, 2001 (see Table 1),
2527 daily quotations of the Dow Jones Industrial Average (DJIA) index for the period Jan. 2nd, 1990 Dec.
30th, 1999 (see Fig. 1), and 1561 daily quotations of the
WIG20 Warsaw Stock Exchange index (based on 20 blue
chip stocks from the Polish capital market) for the period
Jan. 2nd, 1995 Mar. 30th, 2001 (see Table 1).

1
ACF
0.8

0.6

0.6

0.4

0.4

0.2

0.2

0
0.2
0

0
10

20
30
40
Time lag (days)

50

0.2
0

10

20
30
40
Time lag (days)

50

1
ACF

ACF

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0
0.2
0

in Finance, Wiley, 2000.


[13] R.C. Blattberg, N. Gonedes, J. Business 47 (1974) 244.
[14] E. Eberlein, U. Keller, Bernoulli 1 (1995) 281.
[15] J.R. Calderon-Rossel, M. Ben-Horim, J. Intern. Business
Studies 13 (1982) 99.
[16] R. Mantegna, H.E. Stanley, Phys. Rev. Lett. 73 (1994)
2946.
[17] U.A. M
uller, M.M. Dacorogna, R.B. Olsen, O.V. Pictet,
M. Schwarz, C. Morgenegg, J. Banking & Finance 14
(1990) 1189.
[18] D.M. Guillaume, M.M. Dacorogna, R.R. Dave, U.A.
M
uller, R.B. Olsen, O.V. Pictet, Finance Stochast. 1
(1997) 95.
[19] A. Weron, R. Weron, Financial Engineering: Derivatives
Pricing, Computer Simulations, Market Statistics, (in
Polish), WNT, Warsaw, 1998.
[20] R. Weron, Physica A 285 (2000) 127.
[21] L. Bachelier, Theorie de la speculation, Annales Scientiques de lEcole Normale Superieure III-17 (1900) 21.
[22] J. Geweke, S. Porter-Hudak, J. Time Series Analysis 4
(1983) 221.
[23] R. Weron, cond-mat/0103510 (2001).
[24] H.E. Hurst, Trans. Am. Soc. Civil Engineers 116 (1951)
770.
[25] B.B. Mandelbrot, J.R. Wallis, Water Resources Res. 5
(1969) 967.
[26] C.-K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E.
Stanley, A.L. Goldberger, Phys. Rev. E 49 (1994) 1684.
[27] P. Bak, M. Paczuski, M. Shubik, Physica A 246 (1997)
430.

ACF

0.8

0
10

20
30
40
Time lag (days)

50

0.2
0

10

20
30
40
Time lag (days)

50

FIG. 4. Lagged autocorrelation functions of rt and


USD/DEM returns, respectively (top panels). Lagged autocorrelation functions of absolute value for the same data sets
(bottom panels). Dashed horizontal lines represent the 95%
condence interval of a Gaussian random walk.

The presented empirical analysis clearly shows that


three simple rules of our model lead to a fat-tailed distribution of returns, long-term dependence in volatility
and no dependence in returns themselves as observed for
market data. Thus we may conclude that this simple
model is a good first approximation of a number of real
financial markets.
We gratefully acknowledge critical comments of an
anonymous referee, which led to a substantial improvement of the paper.

TABLE I. Estimates of the Hurst exponent H for simulated and market data.

Data
Simulation
USD/DEM
JPY/USD
DJIA
WIG20

[1] D. Derrida, P.G. Higgs, J. Phys. A 24 (1991) L985.


[2] F. Schweitzer, J.A. Holyst, Eur. Phys. J. B 15 (2000) 723.
[3] R. Savit, R. Manuca, R. Riolo, Phys. Rev. Lett. 82 (1999)
2203.
[4] A. Cavagna, J.P. Garrahan, I. Giardina, D. Sherrington,
Phys. Rev. Lett. 83 (1999) 4429.
[5] D. Chowdhury , D. Stauer, Eur. Phys. J. B 8 (1999)
477.
[6] R. Cont and J. P. Bouchaud, Macroeconomic Dyn. 4
(2000) 170.
[7] D. Challet, M. Marsili, R. Zecchina, Phys. Rev. Lett. 84
(2000) 1824.
[8] V.M. Eguiluz, M.G. Zimmermann, Phys. Rev. Lett. 85
(2000) 5659.
[9] K. Sznajd-Weron, J. Sznajd, Int. J. Mod. Phys. C 11,
No. 6 (2000).
[10] D. Stauer, A.O. Sousa, S. Moss de Oliveira , Int. J.
Mod. Phys. C 11, No. 6 (2000).
[11] J. Adler, Physica A 171, 453 (1991).
[12] See eg. S. Rachev, S. Mittnik, Stable Paretian Models

Simulation
USD/DEM
JPY/USD
DJIA
WIG20

Method
DFA
GPH
Returns
0.5270
0.4666
0.3653
0.5127
0.5115
0.6154
0.5353
0.5303
0.5790
0.4585
0.4195
0.3560
0.5030
0.4981
0.4604
Absolute value of returns
0.8940
0.9335
0.8931

0.7751
0.8406
0.8761

0.8576
0.9529
0.9287

0.7838
0.9080
0.8357
0.9103
0.9494
0.8262
R/S-AL

,
and denote signicance at the (two-sided) 90%, 95%
and 99% level, respectively. For the R/S-AL and DFA statistics inference is based on empirical Monte Carlo results of
Weron [23], whereas for the GPH statistics on asymptotic
distribution of the estimate of H [22].